Rong Xu [Wed, 16 Nov 2016 20:50:06 +0000 (20:50 +0000)]
Make block placement deterministic
We fail to produce bit-to-bit matching stage2 and stage3 compiler in PGO
bootstrap build. The reason is because LoopBlockSet is of SmallPtrSet type
whose iterating order depends on the pointer value.
This patch fixes this issue by changing to use SmallSetVector.
Geoff Berry [Wed, 16 Nov 2016 19:35:19 +0000 (19:35 +0000)]
[AArch64] Handle vector types in replaceZeroVectorStore.
Summary:
Extend replaceZeroVectorStore to handle more vector type stores,
floating point zero vectors and set alignment more accurately on split
stores.
Tom Stellard [Wed, 16 Nov 2016 18:42:17 +0000 (18:42 +0000)]
AMDGPU/SI: Avoid creating unnecessary copies in the SIFixSGPRCopies pass
Summary:
1. Don't try to copy values to and from the same register class.
2. Replace copies with of registers with immediate values with v_mov/s_mov
instructions.
The main purpose of this change is to make MachineSink do a better job of
determining when it is beneficial to split a critical edge, since the pass
assumes that copies will become move instructions.
This prevents a regression in uniform-cfg.ll if we enable critical edge
splitting for AMDGPU.
Sanjay Patel [Wed, 16 Nov 2016 17:42:40 +0000 (17:42 +0000)]
[x86] add fake scalar FP logic instructions to ReplaceableInstrs to save some bytes
We can replace "scalar" FP-bitwise-logic with other forms of bitwise-logic instructions.
Scalar SSE/AVX FP-logic instructions only exist in your imagination and/or the bowels of
compilers, but logically equivalent int, float, and double variants of bitwise-logic
instructions are reality in x86, and the float variant may be a shorter instruction
depending on which flavor (SSE or AVX) of vector ISA you have...so just prefer float all
the time.
This is a preliminary step towards solving PR6137:
https://llvm.org/bugs/show_bug.cgi?id=6137
Lang Hames [Wed, 16 Nov 2016 17:31:09 +0000 (17:31 +0000)]
[Orc] Re-enable the RPC unit test disabled in r286917.
This unit test infinite-looped on s390x due to a thread_yield being optimized
out. I've updated the QueueChannel class (where thread_yield was called) to use
a condition variable instead. This should cause the unit test to behave
correctly.
Simon Pilgrim [Wed, 16 Nov 2016 14:48:32 +0000 (14:48 +0000)]
[X86][AVX512] Autoupgrade lossless i32/u32 to f64 conversion intrinsics with generic IR
Both the (V)CVTDQ2PD (i32 to f64) and (V)CVTUDQ2PD (u32 to f64) conversion instructions are lossless and can be safely represented as generic SINT_TO_FP/UINT_TO_FP calls instead of x86 intrinsics without affecting final codegen.
Simon Dardis [Wed, 16 Nov 2016 11:29:07 +0000 (11:29 +0000)]
[mips] Fix unsigned/signed type error
MipsFastISel uses a a class to represent addresses with a signed member
to represent the offset. MipsFastISel::emitStore, emitLoad and computeAddress
all treated the offset as being positive. In cases where the offset was
actually negative and a frame pointer was used, this would cause the constant
synthesis routine to crash as it would generate an unexpected instruction
sequence when frame indexes are replaced.
Craig Topper [Wed, 16 Nov 2016 05:24:10 +0000 (05:24 +0000)]
[X86] Remove the scalar intrinsics for fadd/fsub/fdiv/fmul
Summary: These intrinsics have been unused for clang for a while. This patch removes them. We auto upgrade them to extractelements, a scalar operation and then an insertelement. This matches the sequence used by clangs intrinsic file.
Davide Italiano [Wed, 16 Nov 2016 05:10:28 +0000 (05:10 +0000)]
[ELF] Convert ELF.h to Expected<T>.
This has two advantages:
1) We slowly move away from ErrorOr to the new handling interface,
in the hope of having an uniform error handling in LLVM, eventually.
2) We're starting to have *meaningful* error messages for invalid
object ELF files, rather than a generic "parse error". At some point
we should include also the offset to improve the quality of the
diagnostic.
[XRay][docs] Define requirements on installed log handlers.
Summary:
We update the documentation to define what the requirements are for the
provided XRay log handler. This is to make it clear that the function
pointer provided must do internal synchronisation and that there are no
guarantees provided by XRay on when the function shall be invoked once
it has been installed as a log handler.
Quentin Colombet [Wed, 16 Nov 2016 01:07:12 +0000 (01:07 +0000)]
[RegAllocGreedy] Record missed hint for late recoloring.
In https://reviews.llvm.org/D25347, Geoff noticed that we still have
useless copy that we can eliminate after register allocation. At the
time the allocation is chosen for those copies, they are not useless
but, because of changes in the surrounding code, later on they might
become useless.
The Greedy allocator already has a mechanism to deal with such cases
with a late recoloring. However, we missed to record the some of the
missed hints.
Fixed the lost FastMathFlags for CALL operations in SLPVectorizer.
Reviewer: Michael Zolotukhin.
Differential Revision: https://reviews.llvm.org/D26575
Justin Lebar [Wed, 16 Nov 2016 00:44:47 +0000 (00:44 +0000)]
[BypassSlowDivision] Handle division by constant numerators better.
Summary:
We don't do BypassSlowDivision when the denominator is a constant, but
we do do it when the numerator is a constant.
This patch makes two related changes to BypassSlowDivision when the
numerator is a constant:
* If the numerator is too large to fit into the bypass width, don't
bypass slow division (because we'll never run the smaller-width
code).
* If we bypass slow division where the numerator is a constant, don't
OR together the numerator and denominator when determining whether
both operands fit within the bypass width. We need to check only the
denominator.
Rui Ueyama [Wed, 16 Nov 2016 00:38:33 +0000 (00:38 +0000)]
Fix Modi and File count if there are more than 65535 modules/files.
These numbers are intended to be capped at 65535, but
`std::max<uint16_t>(UINT16_MAX, N)` always returns N for any N because
the expression is the same as `std::max((uint16_t)UINT16_MAX, (uint16_t)N)`.
Always use relative jump table encodings on PowerPC64.
For the default, small and medium code model, use the existing
difference from the jump table towards the label. For all other code
models, setup the picbase and use the difference between the picbase and
the block address.
Overall, this results in smaller data tables at the expensive of one or
two more arithmetic operation at the jump site. Given that we only create
jump tables with a lot more than two entries, it is a net win in size.
For larger code models the assumption remains that individual functions
are no larger than 2GB.
Kevin Enderby [Tue, 15 Nov 2016 23:07:41 +0000 (23:07 +0000)]
General clean up of Mach-O error handling in llvm-objdump.
To get a good error message for all files that could contain Mach-O
files the code in llvm-objdump needs to use the archive member name
and name of the architecture of a slice of a universal file in those cases
where the error come from a Mach-O file in an archive or a universal file.
Most of this is fixed by moving the call to checkSymbolTable() into
ProcessMachO() and calling it when the operation needs the symbol
table. And then calling the form of report_error() that has the
ArchiveName and ArchitectureName arguments. One other place
needed to call this form of report_error() also with these arguments.
Also changed the code in MachODump.cpp to not use report_fatal_error()
and use report_error() instead to make the code smaller and cleaner. All
cases of this are for errors with the symbol table which should now never
be tripped since checkSymbolTable() should be called first to get a good
error message in these cases.
Kuba Brecka [Tue, 15 Nov 2016 21:07:03 +0000 (21:07 +0000)]
Fix llvm-symbolizer to correctly sort a symbol array and calculate symbol sizes
Sometimes, llvm-symbolizer gives wrong results due to incorrect sizes of some symbols. The reason for that was an incorrectly sorted array in computeSymbolSizes. The comparison function used subtraction of unsigned types, which is incorrect. Let's change this to return explicit -1 or 1.
The wave barrier represents the discardable barrier. Its main purpose is to
carry convergent attribute, thus preventing illegal CFG optimizations. All lanes
in a wave come to convergence point simultaneously with SIMT, thus no special
instruction is needed in the ISA. The barrier is discarded during code generation.
Sanjay Patel [Tue, 15 Nov 2016 18:44:53 +0000 (18:44 +0000)]
[x86] auto-generate checks; NFC
Also, fix the test params to use an attribute rather than a CPU model
and remove the AVX run because that does nothing but check for a 'v'
prefix in all of these tests.
Wei Mi [Tue, 15 Nov 2016 18:35:53 +0000 (18:35 +0000)]
[LSR] Allow formula containing Reg for SCEVAddRecExpr related with outerloop.
In RateRegister of existing LSR, if a formula contains a Reg which is a SCEVAddRecExpr,
and this SCEVAddRecExpr's loop is an outerloop, the formula will be marked as Loser
and dropped.
Suppose we have an IR that %for.body is outerloop and %for.body2 is innerloop. LSR only
handle inner loop now so only %for.body2 will be handled.
Using the logic above, formula like
reg(%array) + reg({1,+, %size}<%for.body>) + 1*reg({0,+,1}<%for.body2>) will be dropped
no matter what because reg({1,+, %size}<%for.body>) is a SCEVAddRecExpr type reg related
with outerloop. Only formula like
reg(%array) + 1*reg({{1,+, %size}<%for.body>,+,1}<nuw><nsw><%for.body2>) will be kept
because the SCEVAddRecExpr related with outerloop is folded into the initial value of the
SCEVAddRecExpr related with current loop.
But in some cases, we do need to share the basic induction variable
reg{0 ,+, 1}<%for.body2> among LSR Uses to reduce the final total number of induction
variables used by LSR, so we don't want to drop the formula like
reg(%array) + reg({1,+, %size}<%for.body>) + 1*reg({0,+,1}<%for.body2>) unconditionally.
From the existing comment, it tries to avoid considering multiple level loops at the same time.
However, existing LSR only handles innermost loop, so for any SCEVAddRecExpr with a loop other
than current loop, it is an invariant and will be simple to handle, and the formula doesn't have
to be dropped.
Wei Mi [Tue, 15 Nov 2016 17:34:52 +0000 (17:34 +0000)]
[IndVars] Change the order to compute WidenAddRec in widenIVUse.
When both WidenIV::getWideRecurrence and WidenIV::getExtendedOperandRecurrence
return non-null but different WideAddRec, if getWideRecurrence is called
before getExtendedOperandRecurrence, we won't bother to call
getExtendedOperandRecurrence again. But As we know it is possible that after
SCEV folding, we cannot prove the legality using the SCEVAddRecExpr returned
by getWideRecurrence. Meanwhile if getExtendedOperandRecurrence returns non-null
WideAddRec, we know for sure that it is legal to do widening for current instruction.
So it is better to put getExtendedOperandRecurrence before getWideRecurrence, which
will increase the chance of successful widening.
Simon Pilgrim [Tue, 15 Nov 2016 16:24:40 +0000 (16:24 +0000)]
[X86][SSE] Improve SINT_TO_FP of boolean vector results (signum)
This patch helps avoids poor legalization of boolean vector results (e.g. 8f32 -> 8i1 -> 8i16) that feed into SINT_TO_FP by inserting an early SIGN_EXTEND and so help improve the truncation logic.
This is not necessary for AVX512 targets where boolean vectors are legal - AVX512 manages to lower ( sint_to_fp vXi1 ) into some form of ( select mask, 1.0f , 0.0f ) in most cases.
Diana Picus [Tue, 15 Nov 2016 15:38:15 +0000 (15:38 +0000)]
[ARM] Make sure GlobalISel is only initialized once. NFCI
Move some code inside the proper 'if' block to make sure it is only run once,
when the subtarget is first created. Things can still break if we use different
ARM target machines or if we have functions with different 'target-cpu' or
'target-features', we should fix that too in the future.
Robert Lougher [Tue, 15 Nov 2016 14:27:33 +0000 (14:27 +0000)]
[LoopVectorizer] When estimating reg usage, unused insts may "end" another use
The register usage algorithm incorrectly treats instructions whose value is
not used within the loop (e.g. those that do not produce a value).
The algorithm first calculates the usages within the loop. It iterates over
the instructions in order, and records at which instruction index each use
ends (in fact, they're actually recorded against the next index, as this is
when we want to delete them from the open intervals).
The algorithm then iterates over the instructions again, adding each
instruction in turn to a list of open intervals. Instructions are then
removed from the list of open intervals when they occur in the list of uses
ended at the current index.
The problem is, instructions which are not used in the loop are skipped.
However, although they aren't used, the last use of a value may have been
recorded against that instruction index. In this case, the use is not deleted
from the open intervals, which may then bump up the estimated register usage.
This patch fixes the issue by simply moving the "is used" check after the loop
which erases the uses at the current index.
Tony Jiang [Tue, 15 Nov 2016 14:25:56 +0000 (14:25 +0000)]
[PowerPC] Implement BE VSX load/store builtins - llvm portion.
This patch implements all the overloads for vec_xl_be and vec_xst_be. On BE,
they behaves exactly the same with vec_xl and vec_xst, therefore they are
simply implemented by defining a matching macro. On LE, they are implemented
by defining new builtins and intrinsics. For int/float/long long/double, it
is just a load (lxvw4x/lxvd2x) or store(stxvw4x/stxvd2x). For char/char/short,
we also need some extra shuffling before or after call the builtins to get the
desired BE order. For int128, simply call vec_xl or vec_xst.
Diana Picus [Tue, 15 Nov 2016 14:11:11 +0000 (14:11 +0000)]
Get GlobalISel to build on Linux after r286407
r286407 has introduced calls to llvm::AddLandingPadInfo, which lives in the
SelectionDAG component. Add it to LLVMBuild to avoid linker failures on Linux.
Zvi Rackover [Tue, 15 Nov 2016 13:29:23 +0000 (13:29 +0000)]
[X86][FastISel] Fix lowering of overflow result on AVX512 targets
Summary:
Fix a case where the overflow value of type i1, which is legal on AVX512, was assigned to a VK1 register class.
We always want this value to be assigned to a GPR since the overflow return value is lowered to a SETO instruction.
Introduce TLI predicative for base-relative Jump Tables.
For 64bit ABIs it is common practice to use relative Jump Tables with
potentially different relocation bases. As the logic for the jump table
itself doesn't depend on the relocation base, make it easier for targets
to use the generic logic. Start by dropping the now redundant MIPS logic.
Daniel Sanders [Tue, 15 Nov 2016 09:51:02 +0000 (09:51 +0000)]
[tablegen] Extract portions of AsmMatcherEmitter for re-use by another generator. NFC.
Summary:
This change is preparation for a change that will allow targets to verify that the instructions
they emit meet the predicates they specify. This is useful to ensure that C++
legalization/lowering/instruction-selection doesn't incorrectly select code for a different
subtarget than intended. Such cases are not caught by the integrated assembler when emitting
instructions directly to an object file.
Adam Nemet [Tue, 15 Nov 2016 08:40:51 +0000 (08:40 +0000)]
[opt-viewer] Add support for libYAML for faster parsing
This results in a speed-up of over 6x on sqlite3.
Before:
$ time -p /org/llvm/utils/opt-viewer/opt-viewer.py ./MultiSource/Applications/sqlite3/CMakeFiles/sqlite3.dir/sqlite3.c.opt.yaml html
real 415.07
user 410.00
sys 4.66
After with libYAML:
$ time -p /org/llvm/utils/opt-viewer/opt-viewer.py ./MultiSource/Applications/sqlite3/CMakeFiles/sqlite3.dir/sqlite3.c.opt.yaml html
real 63.96
user 60.03
sys 3.67
I followed these steps to get libYAML working with PyYAML: http://rmcgibbo.github.io/blog/2013/05/23/faster-yaml-parsing-with-libyaml/
Zvi Rackover [Tue, 15 Nov 2016 06:34:33 +0000 (06:34 +0000)]
[X86][GlobalISel] Add minimal call lowering support to the IRTranslator
Summary:
Add basic functionality to support call lowering for X86.
Currently only supports functions which return void and take zero arguments.
Inspired by commit 286573.
Greg Clayton [Tue, 15 Nov 2016 01:23:06 +0000 (01:23 +0000)]
Improve DWARF parsing speed by improving DWARFAbbreviationDeclaration
This patch gets a DWARF parsing speed improvement by having DWARFAbbreviationDeclaration instances know if they have a fixed byte size. If an abbreviation has a fixed byte size that can be calculated given a DWARFUnit, then parsing a DIE becomes two steps: parse ULEB128 abbrev code, and then add constant size to the offset.
This patch also adds a fixed byte size to each DWARFAbbreviationDeclaration::AttributeSpec so that attributes can quickly skip their values if needed without the need to lookup the fixed for size.
Notable improvements:
- DWARFAbbreviationDeclaration::findAttributeIndex() now returns an Optional<uint32_t> instead of a uint32_t and we no longer have to look for the magic -1U return value
- Optional<uint32_t> DWARFAbbreviationDeclaration::findAttributeIndex(dwarf::Attribute attr) const;
- DWARFAbbreviationDeclaration now has a getAttributeValue() function that extracts an attribute value given a DIE offset that takes advantage of the DWARFAbbreviationDeclaration::AttributeSpec::ByteSize
- bool DWARFAbbreviationDeclaration::getAttributeValue(const uint32_t DIEOffset, const dwarf::Attribute Attr, const DWARFUnit &U, DWARFFormValue &FormValue) const;
- A DWARFAbbreviationDeclaration instance can return a fixed byte size for itself so DWARF parsing is faster:
- Optional<size_t> DWARFAbbreviationDeclaration::getFixedAttributesByteSize(const DWARFUnit &U) const;
- Any functions that used to take a "const DWARFUnit *U" that would crash if U was NULL now take a "const DWARFUnit &U" and are only called with a valid DWARFUnit
projects / llvm / log